Active Gaussian Splatting for Autonomous Reconstruction Using D-Optimal Next Best View Planning

Abstract

Autonomous 3D reconstruction of geological features is a critical capability for planetary surface exploration, where communication latency prohibits manual view selection. Conventional approaches survey the environment exhaustively before selecting informative viewpoints, incurring unnecessary observation overhead. We propose an active Gaussian Splatting framework in which a drone autonomously selects the Next Best View (NBV) to maximise reconstruction quality of a target rock. Our approach assumes that an initial survey yields a high-fidelity point cloud of the target, from which a mesh model is derived and used to initialise 100,000 3D Gaussians on the surface. Gaussian positions are frozen at mesh-surface locations, and only appearance parameters (colour, scale, rotation, and opacity) are learned online from incrementally acquired images. Four evenly-spaced seed images initialise the model, after which, at each step, a diagonal Fisher Information Hessian is accumulated over the current training set and the candidate viewpoint that maximises the D-optimal information gain is selected as the next observation. We further provide a theoretical justification showing that convergence between steps is a mathematical necessity for the Taylor linearisation underlying the Fisher criterion, not merely an engineering heuristic.